Prompt injection via email can hijack your AI agent — jev-guard adds risk-scored guards
sujingshen · x · 2026-10-18
Giving Claude your Gmail, Drive and a browser with no supervision is dangerous: one poisoned email or shady web page (or an unread GitHub skill) can push your agent into running rm -rf or exfiltrating files.
The quoted thread offers a concrete guardrail workflow with jev-guard:
- npm i -g jev-guard
- Configure a key (Vercel AI Gateway and OpenRouter keys work too)
- jev-guard key "your-key"
- jev-guard install claude wires it into Claude Code's hooks
- Test: jev-guard check Bash '{"command":"rm -rf /"}' → DENY, risk 3.0/3
After setup, every tool call gets a risk score and high-risk commands are denied.
More from coding & agent
- Open-weight TYGER speech decision model makes 80ms turn-taking calls for full-duplex voice agents — solyarisoftware · 2026-10-18
- Devs: Opus 5.5 on medium beats Opus on max — eyishazyer · 2026-10-18
- Cursor Super + Claude Pro + OpenCode: can this stack survive heavy autonomous coding? — DarioAmodeiWasTaken · 2026-10-18
- Coding AI wars: model is faster and smarter but 'dogshit at actual engineering' — MarcJSchmidt · 2026-10-18
- Anthropic borrows lawyers' ethics for Claude agents that closed 186 deals with no human approval — rohanpaul_ai · 2026-10-18
- 14-Agent AI Film Pipeline on One RTX 5090: Fixing Cut Defects Between Clips — developervkmp · 2026-10-18